Different learning rate from the same module

How can I assign forward-adaptive learning rates to modules in sequential?

I know in normal case, we can use optimizer and make this like "params" : model.a.parameters(), 'lr' : lr / 100, but I wanna use different learning-rate for loss1 and loss2 which went through the same module.

class X(nn.Module):
   def __init__(self):
      self.a = nn.Conv2d(1,1,1)
      self.b = nn.Conv2d(1,1,1)

   def forward(self, input):
       return self.b(self.a(input))

model = X()
output1 = model(input1)
output2 = model(input2)

loss1 = F.mse_loss(output1, label1)
loss2 = F.mse_loss(output2, label2)

And for loss1, I want both a and b parameters step-up with lr, but for loss2, a for lr and b for -lr.

module a b
loss1 lr lr
loss2 lr -lr

because loss1, loss2 backward through both a and b modules, we can’t make this just by (- loss1).backward() or like that.

can I assign some options when model forwards?


The loss calculation is independent from the optimization step (parameter update) in general.
You could create two optimizers with the specified learning rates and apply them sequentially, if that would fit your use case:

loss1.backward(retain_graph=True) # calculate gradients using loss1
optimizer1.step() # update parameters with lr, lr

loss2.backward() # calculate gradients using loss2
optimizer2.step() # update parameters with lr, -lr

Would this work for you?


Great! It seems to work well!
Thank you for your quick and kind reply!